3D Bounding Box Vehicle Detection via Neural Network Regression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current vehicle detection systems face challenges in accurately and efficiently detecting and tracking vehicles using 2D bounding boxes, which are computationally expensive and do not provide sufficient information about vehicle orientation, while 3D bounding boxes offer better accuracy but at a higher computational cost.

Innovation Solution

A method utilizing a convolutional neural network to directly output three-dimensional bounding boxes, defined by a combination of two quadrilaterals representing the rear and side of detected vehicles, which are computationally efficient and provide orientation information, allowing for accurate vehicle detection and tracking.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If 2D bounding boxes are used for vehicle detection, then computational cost is reduced, but orientation information is insufficient and detection accuracy deteriorates

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidvehicle orientation information
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent transitions from 2D bounding boxes to 3D bounding boxes by adding depth dimension information. The neural network directly outputs 3D bounding box parameters (x, y, z coordinates, width, height, depth, and rotation angles) from 2D image inputs, enabling the system to obtain orientation information while maintaining computational efficiency through direct regression rather than complex 3D reconstruction processes.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If 3D bounding boxes are used for vehicle detection, then orientation information and detection accuracy are improved, but computational cost increases

Engineering Contradiction:
Improvevehicle detection accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent performs preliminary action by training the neural network offline to directly regress 3D bounding box parameters from 2D images. During online detection, the pre-trained network quickly outputs 3D parameters without requiring computationally intensive post-processing or multiple view synthesis, thus achieving real-time performance with accurate 3D vehicle detection.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent extracts only the essential 3D bounding box parameters (position, size, orientation) directly from 2D image features using a streamlined neural network architecture. This extraction approach avoids the computational burden of full 3D object reconstruction while retaining the critical orientation and spatial information needed for autonomous driving applications.

Inventive Principle:
Principle #2Taking out (Extraction)

3Measurement precision

If complex neural network architectures are used to improve detection accuracy, then measurement precision improves, but device complexity and computational burden increase

Engineering Contradiction:
Improvevehicle detection precisionVSAvoidneural network complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the vehicle detection task into distinct components handled by different parts of the neural network: feature extraction layers process 2D image features, while regression layers directly predict 3D bounding box parameters. This segmentation allows the network to focus on specific sub-tasks, improving accuracy without requiring an overly complex monolithic architecture.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3607489B1Direct vehicle detection as 3D bounding boxes using neural network image processing
Publication Date: 2023.05.24 ROBERT BOSCH GMBH
  • EP3607489B1 patent drawingFigure 1~2
  • EP3607489B1 patent drawingFigure 3~4
  • EP3607489B1 patent drawingFigure 5~6

AI summary

Systems and methods of detecting and tracking one or more vehicles in a field of view of an imaging system using neural network processing. An electronic controller receives an input image from a camera mounted on the host vehicle. The electronic controller applies a neural network configured to output a definition of a three-dimensional bounding box based at least in part on the input image. The three-dimensional bounding box indicates a size and a position of a detected vehicle in a field of view of the input image. The three-dimensional bounding box includes a first quadrilateral shape outlining a rear or front of the detected vehicle and a second quadrilateral shape outline a side of the detected vehicle.